NASA Technical Reports Server (NTRS) 20110023793: Surrogate Modeling of High-Fidelity Fracture Simulations for Real-Time Residual Strength Predictions

4.9 MB·English

by  NASA Technical Reports Server (NTRS)

Surrogate Modeling of High-Fidelity Fracture
Simulations for Real-time Residual Strength
Predictions
1 2 3 4
Ashley D. Spear , Amanda R. Priest , Michael G. Veilleux and Anthony R. Ingra(cid:27)ea
Cornell University, Ithaca, NY 14853
5
Jacob D. Hochhalter
NASA Langley Research Center, Hampton, VA 23681
A surrogate model methodology is described for predicting in real time the residual
strength of (cid:29)ight structures with discrete-source damage. Starting with design of ex-
periment, an arti(cid:28)cial neural network is developed that takes as input discrete-source
damage parameters and outputs a prediction of the structural residual strength. Tar-
get residual strength values used to train the arti(cid:28)cial neural network are derived from
3D (cid:28)nite element-based fracture simulations. A residual strength test of a metal-
lic, integrally-sti(cid:27)ened panel is simulated to show that crack growth and residual
strength are determined more accurately in discrete-source damage cases by using
an elastic-plastic fracture framework rather than a linear-elastic fracture mechanics-
based method. Improving accuracy of the residual strength training data would, in
turn, improve accuracy of the surrogate model. When combined, the surrogate model
methodology and high-(cid:28)delity fracture simulation framework provide useful tools for
adaptive (cid:29)ight technology.
1 GraduateStudent,SchoolofCivilandEnvironmentalEngineering,640RhodesHall,AIAAStudentMember.
2 UndergraduateStudent,SchoolofCivilandEnvironmentalEngineering,640RhodesHall.
3 GraduateStudent,SchoolofCivilandEnvironmentalEngineering,640RhodesHall.
4 DwightC.BaumProfessorofEngineering,SchoolofCivilandEnvironmentalEngineering,643RhodesHall.
5 ResearchMaterialsEngineer,Durability&DamageToleranceBranch,MS188E,AIAAMember.
1
Nomenclature
E = elastic modulus (GPa)
(cid:23) = Poisson’s ratio (mm/mm)
(cid:27) = yield stress (MPa)
y
STRI65 = quadratic triangular shell element in ABAQUS [1]
S8R = quadratic reduced-integration shell element in ABAQUS [1]
C3D10 = quadratic tetrahedral elements in ABAQUS [1]
C3D15 = quadratic wedge element in ABAQUS [1]
C3D20(R) = quadratic brick element (reduced-integration) in ABAQUS [1]
a = crack length (cm)
n = number of cracks in discrete-source
(cid:18) = orientation of discrete-source damage, angle between positive x axis and nearest crack
d = distance from middle sti(cid:27)ener to center of discrete-source damage (cm)
x
√
K ;K ;K = mode I, II, and III plane strain stress intensity factors (MPa m)
I II III
√
K ;K = plane strain fracture toughness for modes I and II (MPa m)
Ic IIc
P = damage-dependent allowable traction, residual strength (MPa)
max
P = applied traction (MPa)
MSE = mean squared error as de(cid:28)ned by Eq. (2)
c = correlation coe(cid:30)cient as de(cid:28)ned by Eq. (3)
v
CTD = magnitude of relative displacement between upper and lower fracture surfaces (mm)
CTD = critical value of CTD (mm)
crit
CTD ;CTD ;CTD =opening,in-planesliding,out-of-planeshearingcomponentsofCTD
I II III
(mm)
d = (cid:28)xed characteristic distance behind crack front where CTD is monitored (mm)
da = crack extension (mm)
(n) = script to denote mesh at nth crack increment
(n+1) = script to denote mesh at (n+1)th crack increment
2
I. Introduction
Resilient aircraft control involves adaptive responses to o(cid:27)-nominal (cid:29)ight conditions, including
theincurrenceofstructuraldiscrete-sourcedamageduring(cid:29)ight. Discrete-sourcedamage istypically
manifestedasaresultofastructuralimpactevent,includinghail-andbirdstrike. In2003,anAirbus
A300 operated by DHL was struck by a surface-to-air missile after takeo(cid:27) from Baghdad, Iraq,
causing discrete-source damage to crucial control surfaces of the left wing [2]. In 2008, a Boeing
747-438 operated by Qantas Airways incurred in-(cid:29)ight structural damage to the fuselage and right
wing leading edge following the failure of an onboard oxygen cylinder [3]. Although the aircraft
landed safely in both cases, these examples motivate a need for more resilient, adaptive control
system responses.
In these types of cases, problems associated with in-(cid:29)ight discrete-source damage, for example
inability to sustain original design loads, can be exacerbated by crack propagation from damaged
regions. To avoid unstable crack propagation, load levels must be maintained below a reduced
load-carrying capacity, or residual strength, of damaged (cid:29)ight structures. Adaptive control system
responses might include automatic adjustment of certain (cid:29)ight parameters (e.g. velocity, maximum
acceleration) to accommodate structural residual strength. This accommodation implies that accu-
rate residual strength predictions of (cid:29)ight structures with complex damage con(cid:28)gurations be made
in real time, during (cid:29)ight; this capability currently does not exist for commercial aviation.
Challengestodevelopinganadaptiveresponsetechnologyincludeaccuratelypredictingresidual
strength of discrete-source damaged structures both o(cid:31)ine (i.e. during control system design) and
online (i.e. in real time onboard the aircraft). In the o(cid:31)ine context, researchers have developed
various tools for determining residual strength of thin, damaged metallic structures using elastic-
plastic fracture mechanics (EPFM)-based numerical methods. For example, two common (cid:28)nite
element (FE) modeling techniques involve nodal release and adaptive remeshing. Both techniques
representcracksgeometrically[4]. Theformer,however,prescribespossiblecracktrajectories,which
introduces inherent mesh dependencies into fracture simulations and limits generality of crack path
predictions. Nodal release techniques have been used in 2D [5(cid:21)12] and in 3D [13(cid:21)15] for studying
crack growth parameters and predicting residual strength of structures where the crack path was
3
known a priori and where mesh re(cid:28)nement along the crack path su(cid:30)ciently characterized growth
increments. Adaptive remeshing techniques avoid such mesh dependencies and enable simulation
of arbitrary crack propagation using evolutionary models or criteria [16(cid:21)20]. Adaptive remeshing
techniques have been implemented in both 2D [21] and in 3D [22, 23]. Of the described techniques,
3D, adaptively remeshed, elastic-plastic tearing simulations provide the most general prediction
capabilities for crack growth and residual strength.
It is infeasible to perform a rigorous and computationally intensive crack growth simulation
within the possible short time span following a discrete-source damage event. Thus, an approxima-
tion, or surrogate model, is needed for making online predictions of residual strength. Queipo et al.
provided a complete description of surrogate modeling development and optimization [24]. With
regard to surrogate construction, they described both parametric (e.g. polynomial regression and
Kriging) and nonparametric (e.g. radial basis functions) approaches. In nonparametric approaches,
a global functional form relating system input to system response is not assumed.
Arti(cid:28)cial neural networks (NNs) are a nonparametric surrogate modeling approach and are
trainedtoinferanonlinearmappingfromsysteminputtosystemresponse,oroutput. Thereaderis
referredto[25]foranextensivemethodologyoverviewofthemostwidelyusedtypesofNN.Di(cid:27)erent
typesofNNshavebeenappliedextensivelyfordamagedetection[26(cid:21)32]and,toamuchlesserextent,
for damage assessment. Ouenes et al. employed a NN methodology to predict fracture indicators
(e.g. density of fractures) in naturally fractured rock reservoirs as a function of various geological
andgeophysicaldata[33]. Pidapartietal.employedaNNtopredictresidualstrengthandcorrosion
rate of aging aircraft panels with collinear multi-site damage by training with experimental results
and validating with both experimental results and analytical solutions [34]. Recently, Mohanty et
al. used a Gaussian process (GP) approach to predict fatigue crack growth in aluminum 2024-T351
specimens by training two distinct models, one presented with experimental load parameters as
input and another presented with piezoelectric sensor signals as input [35]. In that work, Mohanty
etal.usedobservedfatiguecracklengthsandgrowthratesasknownoutputfortrainingeachmodel.
Alternatively,NNscanbetrainedusingresultsfromnumericalexperiments,orsimulations[36].
For example, Sankararaman et al. recently used linear-elastic fracture parameters computed from
4
FE analyses to train a GP model as part of a method to statistically infer equivalent initial (cid:29)aw
size in fatigue applications [37]. High-(cid:28)delity numerical simulations can provide training data when
analytically-andexperimentally-deriveddataarelimiteddueeithertoalackofgenerallyapplicable
analytical solutions or to prohibitive costs of obtaining su(cid:30)cient experimental data.
Thepurposeoftheworkpresentedhereistwo-fold: (1)toillustrateamethodologyforcreating
a surrogate model as a real-time residual strength prediction tool and (2) to describe and validate
numerical tools for making accurate residual strength predictions o(cid:31)ine using fully 3D, elastic-
plastic,FE-basedcrackgrowthsimulations. Thehigh-(cid:28)delity,morecomputationallyexpensivetools
describedin(2)canprovidetrainingdatathat,whencoupledwiththesurrogatemodelmethodology
described in (1), can be used in the design of adaptive response technology.
Consistentwithourtwo-foldpurpose,thispaperisdividedintotwoprimarysections. SectionII
illustrates the methodology for developing a surrogate model (in particular, a NN) that predicts
residualstrengthasafunctionofdiscrete-sourcedamageparameters. Themethodologyisillustrated
using a relatively simple proof-of-concept example. The procedure for gathering training data
is described in IIA and IIB. Because an implementation-ready NN is beyond the scope of this
paper, training data for the proof-of-concept example relies on reduced-order residual strength
approximations. After collecting training data, a simple NN is constructed in IIC by optimizing
certain performance parameters. Finally, a sensitivity study is conducted in IID to understand the
e(cid:27)ect of each damage parameter on predicted residual strength speci(cid:28)cally for the proof-of-concept
structure.
SectionIIIimprovesupono(cid:31)ineresidualstrengthpredictiontoolsusedinSectionIIbysimulat-
ing 3D, elastic-plastic tearing. The tools provide more general crack growth simulation capabilities
and can be used to generate accurate residual strength training data. A relatively large, integrally-
sti(cid:27)ened panel (ISP) that exhibits crack branching is simulated in IIIC to validate the tools.
ResultsanddiscussionsfromtheNNproof-of-conceptexampleandfromtheelastic-plastictear-
ingsimulationareprovidedineachrespectivesection. SectionIVo(cid:27)ersasummaryandconclusions
for the entirety of this work.
5
II. Neural Network Development and Methodology
This section describes the development of a surrogate model for predicting residual strength of
discrete-source damaged aircraft structures in real time. A global functional form is not assumed
for the nonlinear relationship between residual strength and the damage parameters in(cid:29)uencing it;
thus, a nonparametric surrogate model is developed. In particular, a supervised NN is considered
due to rapid prediction capabilities amenable to real-time applications. In Fig. 1, the upper dashed
regionshowsthegeneralizedprocedurefordevelopingaNN(surrogatemodel)thatpredictsresidual
strengthasafunctionofparameterizeddiscrete-sourcedamage. Thelowerdashedregionshowsthe
functionality of the NN (surrogate model) in a real-time context.
Fig. 1 Upper dashed box illustrates a general approach for developing a surrogate model to
predict residual strength of damaged structures. Lower dashed box illustrates how the sur-
rogate model would function onboard an aircraft for predicting residual strength of damaged
structures in real time.
The (cid:28)rst step in this type of surrogate model development is typically referred to as design of
experiment (DOE) [24] and involves obtaining data points that will be used to train and test the
NN. The DOE should be based on the intended application of the NN. For example, if the NN is
intended to provide residual strength predictions in terms of maximum allowable bending moment
in a damaged aircraft wing, then the data points should be gathered using an appropriate wing
structure with applied boundary conditions of interest. Each data point includes sampled input
variable(s) and corresponding known system response(s), called target output. Once the NN has
been trained to map given input to target output, it becomes a useful tool for predicting system
6
response when presented with new input that is within the training range but does not necessarily
correspond to data points used for training.
To illustrate the methodology, a simple NN is developed using a representative wing structure
and reduced-order (linear-elastic) approximations for predicting residual strength. The representa-
tive wing structure is a 61.0 x 91.4 cm2 integrally-sti(cid:27)ened panel (ISP) with three blade sti(cid:27)eners
each 5.1 cm in height, as shown in Fig. 2. The ISP skin and sti(cid:27)eners are 2.3 mm thick. The panel
is modeled as linear-elastic with E = 71:0 GPa and (cid:23) = 0:33, similar to values for a 2XXX series
lower wing skin aluminum alloy (AA).
Fig. 2 Schematic of ISP model with dimensions similar to those used in [38]. Plan view (top)
and cross-section showing integral blade sti(cid:27)eners (bottom). A damage-containing region is
modeled using 3D solid elements (enclosed in shell-solid boundary) while remainder of panel
is modeled with shell elements. All dimensions in cm.
Multiple FE models of the uncracked panel are constructed using ABAQUS⃝R [1]. A shell-
solid modeling technique is employed, where each panel is modeled using 3D solid elements in a
region that will contain damage and shell elements elsewhere, as depicted schematically in Fig. 2.
In this way, 3D constraint is inherently captured along crack fronts using fully 3D solid elements,
while shell elements help maintain a level of computational e(cid:30)ciency yet are able to capture out-
of-plane deformation and possible buckling. The shell and solid element regions are joined using
a coupling constraint, whereby resultant forces and moments acting at shell edge nodes on the
shell-solid boundary are distributed as forces acting at nodes located in a region of in(cid:29)uence on the
7
solidsurfaceoftheshell-solidboundary. Ameshre(cid:28)nementstudyiscarriedouttoensureadequate
discretization of the panel models. Uncracked panels are modeled using approximately 50 STRI65,
2000 S8R, and between 1800 and 17,300 C3D20R elements, depending on the size of the damaged
region. Boundary conditions for the ISP models are de(cid:28)ned to emulate tensile loading conditions
for a region of the lower wing surface and are shown schematically in Fig. 2.
A supplementary study was carried out to determine shell-solid boundary e(cid:27)ects on nearby
crackfrontsinordertominimizethesizeofthesolidregionwithouta(cid:27)ectingstressintensityfactors
(SIFs) computed along nearby crack fronts. Maintaining fracture parameter accuracy is especially
important since fracture parameters are used to predict structural residual strength (described in
IIB),whichisinturnusedtotraintheNN(describedinIIC). Thesupplementarystudyconsidered
a 61.0 x 91.4 cm2 unsti(cid:27)ened panel of the same (linear-elastic) material and thickness as the ISP
described above. The panel had a single, 12.7 cm long, centrally-located through-crack oriented in
thex direction(normaltoappliedtensileload). Bothtensileandbendingconditionswereconsidered
in the study. The panel was modeled entirely with shell elements except for a region containing the
crack, which was modeled with 3D solid elements. All model parameters remained constant while
varying the size (both in-plane dimensions) of the square-shaped solid region, therefore varying the
distance from the shell-solid boundary to the crack front. The size of the solid region was initially
slightlylargerthanthelengthofthecrackandwasincreaseduntilcomputedSIFsconverged. Results
fromthesupplementarystudyindicatedthatforastatic,linear-elasticcrack,thedistancefromshell-
solid boundary to nearest crack front should be no less than 25% of the crack length. This ensures
that the shell-solid boundary has negligible e(cid:27)ect on computed SIFs. The same rule-of-thumb is
applied to the example ISP models described in the NN study.
The following sections describe the generally applicable methodology for developing a NN as a
real-time residual strength prediction tool.
A. Input Variables: Discrete-source Damage Parameters
Discrete-source damage in this work is represented by a symmetric, star-shaped, array of equi-
length cracks, as depicted in Fig. 3(b). This representation of discrete-source damage is motivated
8
by observations of petaling caused by penetration damage to thin metallic structures, see Fig. 3(a).
If all of the cracks in the star-shaped array of Fig. 3(b) separate under load (i.e. there are no crack
closure e(cid:27)ects), then the cracked region transfers no load and e(cid:27)ectively represents a circular hole
withpetalingedges,similartothatshowninFig.3(a). Thedamagerepresentationisparameterized
by the four variables n, a, d , and (cid:18), which are postulated to in(cid:29)uence residual strength of the ISP.
x
Fig. 3 (a) Petaling on the reverse side of a metallic sheet subject to explosive, discrete-source
damage [39]. (b) Schematic showing the representation and parameterization of discrete-
source in the NN example described in this work.
The sample space of damage con(cid:28)gurations is de(cid:28)ned by a range of values for each parameter.
Ranges can be speci(cid:28)ed based on accident reports, photographic evidence, potential structural
threats, design speci(cid:28)cations, and so forth. Inherently, the NN predictions are valid only for input
parameter values within the range of training data. Thus, it is necessary to de(cid:28)ne the sample space
based on the particular NN application. In the example NN, ranges for each damage parameter
are limited to some extent by the ISP geometry. Each range is given in Table 1. The parameter
n, takes integer values ranging from two to six. The range of (cid:18) depends on n due to the de(cid:28)nition
of orientation and the symmetry of the star-shaped con(cid:28)guration. The range of a is de(cid:28)ned in
terms of ISP bay width, from 1=8 ∗ baywidth to 1=4 ∗ baywidth. Due to symmetry of the ISP
9
model, the parameter d ranges from 10.2 cm (damage centered in mid-bay) to 0 (damage centered
x
at middle sti(cid:27)ener). If the damage is located such that the damage-containing, solid FE region
overlaps anywhere with the middle sti(cid:27)ener, the sti(cid:27)ener is assumed to be severed in the damaged
region and is modeled explicitly as such.
Table 1 Range of values associated with each damage parameter in the example NN.
Damage Parameter Range
n 2-6
(cid:18): n=2 (deg) 0-90
(cid:18): n=3 (deg) 0-60
(cid:18): n=4 (deg) 0-45
(cid:18): n=5 (deg) 0-36
(cid:18): n=6 (deg) 0-30
a (cm) 1.27-5.08
d (cm) 0-10.2
x
The damage parameter space is sampled to obtain damage con(cid:28)gurations, each expressed as
a combination of input parameters (n;(cid:18);a;d ). The space of variables can be sampled using a
x
number of di(cid:27)erent sampling methods, including random, strati(cid:28)ed, and Latin Hypercube [40].
Latin Hypercube Sampling (LHS) is a type of strati(cid:28)ed sampling method that guarantees each
partition, or stratum, of input variable space is sampled, though not necessarily uniformly. In this
work,LHSisperformed(cid:28)vetimesforeachofthevariables((cid:18);a;andd ). Eachofthe(cid:28)veLHSruns
x
corresponds to a di(cid:27)erent value of n (two, ..., six cracks) and requires the number of partitions to
be speci(cid:28)ed. The MATLAB⃝R implementation for LHS is used here [41], where output is provided
in the range from zero to one. Each sample value is then scaled to the respective parameter range
according to Table 1.
Table2showsalldamagecon(cid:28)gurations(26intotal)thataremodeledintheISPNNexample,
where each con(cid:28)guration is expressed in terms of sampled input parameters. For each damage
con(cid:28)guration,thex andy dimensionsofthesquare,damage-containing,solidFEregionareprovided
in the sixth column. The x and y dimensions are each 25% larger than the diameter of the star-
shaped damage (i.e. 1:25∗2a), as suggested by the supplementary shell-solid boundary e(cid:27)ect study
described above. The last column speci(cid:28)es whether or not the solid, damaged region severs the
middle sti(cid:27)ener. If so, the portion of the sti(cid:27)ener that intersects the solid model region is removed;
10